The First Giant, and What Its Lens Actually Shows

GLM-4.5-Air is a 106-billion-parameter mixture-of-experts model, 206 GB of bfloat16 weights, too large for any single GPU we own. It is the spectrum’s first frontier-scale model with a full Jacobian-lens row.

How

The weights were never resident in VRAM. The streamed lens materialises one transformer block at a time from striped NVMe, runs the forward-mode dual through it, and returns it to the meta device before loading the next. One weight-read per question, ~5 minutes per readout. A cosine gate against a resident reference model requires 0.999+ agreement before any number is trusted. Three instrument bugs were fixed getting here, each logged in the corrections ledger. The worst never showed on any smaller model. A reverse-mode graph kept block weights pinned after we thought they were freed, harmless until a model too big for VRAM made it fatal.

What it shows, and what it does not

The lens produced 20 readouts, one per banking question, and the token mass is interpretable: mostly British/Australian register (learnt, whilst, upto, maths), fitting the en-AU prompts, plus the model’s bilingual vocabulary.

The direction itself is close to question-independent: 0.98-cosine stable whether the question is about student accounts or term deposits. On the smaller models the lead tokens change with the question; they carry content. Here the lead is register, a stylistic direction sitting on top of whatever is asked.

Three explanations are live; the data cannot yet choose between them. It could be a property of the model (a strong register prior dominating its output geometry), a property of scale (larger models carrying heavier stylistic directions), or an artifact of the instrument (a top-64 window too shallow to reach the content-carrying directions under a dominant register one). A deeper-window follow-up is queued.

Scope

The answers the lens read came from an FP8-quantised serve of the checkpoint, the only way to fit generation on two GPUs; the lens itself read the full bfloat16 weights. The generation is 8-bit, the measurement 16-bit; keep the two apart. The row is an instrument milestone, the first giant lensed white-box, and its interpretation stays narrow until the follow-up runs.